Estimation of Soil Properties Using a Combination of Spectral and Scalar Sensor Data

The measurement of soil properties on site-specific basis is desired for modern production agriculture. This paper addresses this need using an on-the-go in-situ spectrophotometer to acquire NIR reflectance spectra of soil. The spectral data is optionally augmented with electrical conductivity, temperature, and pH sensor data. Calibrations are a particular problem in that they may need to be optimized for particular soils and temporal conditions in order to achieve acceptable accuracy. This work tests the effectiveness of locally weighted partial least squares regression (LWPLS), a recently developed memory-based learning algorithm, in creating calibrations for the measurement of soil properties. As its name implies, the algorithm inherently optimizes predictions based upon the data space near the query point. LWPLS is used to create calibrations for multiple soil properties using two data sets with measurements from a total of 7 fields. In comparisons with three classical regression algorithms, LWPLS is found to produce calibrations with the highest accuracy for the majority of soil properties. None of the algorithms showed a significant advantage in improving calibrations when the spectral data was augmented with scalar sensor data